Expected classpath expansion length to be %zu but instead go

Error message

Expected classpath expansion length to be %zu but instead got %zu

What it means

getClassPath expands wildcards in two passes: one getClassPath_helper call sizes the buffer, a second fills it. If the length returned by the second pass differs from the first, this invariant message is printed, the buffer is freed, and setup fails. The usual cause is the directory contents changing between the two passes, since the sizing pass ran earlier.

Source

Thrown at hadoop-hdfs-project/hadoop-hdfs-native-client/src/main/native/libhdfs/jni_helper.c:629

    }

    // Allocte memory for expanded classpath string
    expandedClasspath = calloc(length, sizeof(char));
    if (expandedClasspath == NULL) {
        fprintf(stderr, "getClassPath: failed calloc: %s\n", strerror(errno));
        return NULL;
    }

    // Actual expansion
    retval = getClassPath_helper(classpath, expandedClasspath);
    if (retval < 0) {
        free(expandedClasspath);
        return NULL;
    }

    // This should not happen, but dotting i's and crossing t's
    if (retval != length) {
        fprintf(stderr,
          "Expected classpath expansion length to be %zu but instead got %zu\n",
          length, retval);
        free(expandedClasspath);
        return NULL;
    }

#ifdef _LIBHDFS_JNI_HELPER_DEBUGGING_ON_
    printf("===============\n");
    printf("Allocated %zd for expanding classpath\n", length);
    printf("Used %zu for expanding classpath\n", strlen(expandedClasspath) + 1);
    printf("Expanded classpath=%s\n", expandedClasspath);
#endif

    return expandedClasspath;
}


/**

View on GitHub (pinned to 2add963021)

Solutions

Retry the operation: with a stable directory the two passes agree
Point wildcard classpath entries at versioned, immutable directories
Pre-expand the CLASSPATH (hadoop classpath --glob) to skip the two-pass logic entirely

Defensive patterns

Strategy: retry

Try / catch

/* transient: directory changed between sizing and fill passes */
for (int attempt = 0; attempt < 3; attempt++) {
    hdfsFS fs = hdfsConnect(uri, user);
    if (fs) break;
    if (errno != EINTERNAL) break;          /* only retry JVM-bootstrap failures */
    /* brief backoff; with a stable directory the two passes agree */
    usleep(200 * 1000);
}

Prevention

Point wildcard classpath entries at immutable, versioned directories
Pre-expand CLASSPATH with `hadoop classpath --glob` to skip the two-pass expansion
Avoid upgrading jars in place while clients start

When it happens

Trigger: Jars added or removed in a wildcard directory between the sizing and filling passes; concurrent deployments writing into the same install tree while a client initializes.

Common situations: Automated upgrade jobs writing into share/hadoop/* while processes start; shared mounts used simultaneously by many clients.

Related errors

AI-assisted analysis of apache/hadoop@2add963021 (2026-08-22). Data as JSON: /api/errors/9515e0e268d9ff54. Report an issue: GitHub.